AskDocsAgentic RAG Knowledge Assistant

A production-style agentic RAG assistant that answers questions over your documents with source citations, and refuses to answer when it can't ground the response.

In numbers

100%: answer relevance & retrieval hit-rate
0.83–1.0: faithfulness (LLM-as-judge)

What I did

1.
Built a production-style agentic RAG assistant providing grounded, source-cited answers over user-uploaded documents, combining ChromaDB vector search, sentence-transformers embeddings, and a hallucination guardrail that refuses ungrounded answers.
2.
Engineered an LLM agent that routes each query to the appropriate tool (document retrieval, web search or calculator) using structured tool-use and deterministic execution.
3.
Designed a custom RAGAS-style evaluation suite measuring faithfulness, answer-relevance, and context hit-rate via LLM-as-judge, achieving 100% answer-relevance and retrieval hit-rate and 0.83–1.0 faithfulness on a labelled QA set.
4.
Shipped it end-to-end: Streamlit web app, multi-format ingestion (PDF, CSV, URL), Docker containerisation, and a provider-agnostic LLM layer running locally on Ollama or deployable to cloud via OpenAI-compatible APIs.

Notes

Most RAG demos stop at "it answered". AskDocs was built around the harder question: can you trust the answer?

Every response is grounded in retrieved chunks and cited back to its source. When the retrieved evidence doesn't support an answer, a guardrail makes the assistant say so rather than guess. An agent layer decides per query whether to search the documents, search the web, or calculate, and the whole thing is measured with an LLM-as-judge evaluation suite rather than vibes.

Built with

Python · LangChain · ChromaDB · sentence-transformers · Streamlit · Ollama · Docker

Next project

FinSight-3B →